a graph-based word sense disambiguation using...

A Graph-based Word Sense Disambiguation Using

Collocation

JungGil Cho

Division of Computer Engineering, SungKyul University

Abstract. Word sense disambiguation (WSD) is an essential part of Natural

Language Processing (NLP) task to identify the senses of words in context. A

new WSD approach based on collocation is proposed in this paper. In this

approach, collocation is used to characterize the senses of words and the graph

connectivity is measured by using various meanings which exist in the

WordNet, when no collocation exists among the words.

Keywords: Collocation, Word Sense Disambiguation (WSD), Tree-based

Association Rule (TAR), WordNet.

1 Introduction

The precise sense of an ambiguous word, which is general feature of all natural

languages, can be selected from the context where words are located, and can be

defined as the task in which inherent sense of polysemy word is allocated

automatically in context. This task is called WSD, which figures out the precise

senses of words in context.

The unsupervised WSD method is used in case of performing WSD without data

learning, and is classified into graph-based ones and the similarity-based ones. In

graph-based methods, a sense graph is built from the words in context, and it is

processed to select the most appropriate meaning for each word from the created

sense graph. Experimental comparisons between the two algorithm types indicate that

graph-based algorithms outperform similarity-based ones, often by a significant

margin [3].

In this paper, a new WSD approach is proposed which shows better performances

over the conventional unsupervised WSD approaches. The suggested method

establishes the framework to measure the best connectivity by increasing efficiency

with the use of collocation before graph formulation. Moreover, the connectivity is

evaluated suing the degree centrality algorithm in which connectivity measurement is

proved as best. When using the proposed graph-based measurement approach

combined with collocation, better accuracy with better precise meaning is resulted

from the performance evaluation over the conventional approaches.

Advanced Science and Technology Letters Vol.86 (Ubiquitous Science and Engineering 2015), pp.77-80

http://dx.doi.org/10.14257/astl.2015.86.16

ISSN: 2287-1233 ASTL Copyright © 2015 SERSC

2 Related Researches

Collocation refers to a group of practical words that habitually go together, whereas

the sense of a word is figured out by accompanying words. In this case words are to

be classified in terms of co-occurrence relation as well as sense. The co-occurrence

relation means the constraints shown in the sense combination relation, which is

called collocation constraint or selection constraint. Therefore, a collocation is a

combination of words that occur more often and naturally than would be expected by

random, and is divided into strong collocation in which closely-coupled words are

used like nearly one word and weak collocation[4].

3 New Criteria of WSD

3.1 WSD Using Collocations

In this paper, WSD algorithms measure sense accuracy based on co-occurrences and

collocation information. In case of using collocation information, the tendency to

have one meaning for one collocation can be utilized since a word with ambiguity

owns one sense in one literature. Hence, if a word with ambiguity includes collocation,

WSD process can be handled more efficiently by having priority to use collocation

information.

Collocation is classified into three types, idiom, phrasal verbs and collocation in

terms of its characteristics. Especially, the syntactic patterns of collocation are

described in details in Table 1.

Table 1. Syntactic Pattern Types of Collocation

Type Grammar Type Example

Type 1 adjective + noun a huge profit

Type 2 noun + noun a pocket calculator

Type 3 verb + adjective +

noun

learn a foreign language

Type 4 verb + adverb live dangerously

Type 5 adverb + verb half understand

Type 6 adverb + adjective completely soaked

Type 7 verb + preposition +

noun

speak through an interpreter

Advanced Science and Technology Letters Vol.86 (Ubiquitous Science and Engineering 2015)

78 Copyright © 2015 SERSC

3.2 Collocation-based WSD Algorithms

The WSD algorithms proposed in this paper are executed basically by the unit of

sentence. Step-wise execution process of the algorithms, shown in Figure 1, is

explained as below. At the 1st step, if a word with ambiguity includes collocation,

WSD process can be handled more efficiently by having priority to use collocation

information. Secondly, a graph G is built by using set G, and then a score of each

vertex is counted. Finally, the sense of the word is determined from the label with the

highest score out of sense of each word.

---------------------------------------------------------------------------------------------------------------

Input: Sequence : W = (𝑤𝑖 |𝑖 = 1 … 𝑁)

Output: Sequence :(L = (𝐿𝑤𝑖|𝑖 = 1 … 𝑁)

1: Check the collocation and if it matches allocate sense

1.1: Analyze morpheme

1.2: Analyze by word class

1.3: Test by category

1.4: Collocation matching

2: If the collocation matches, go to stage 7

3: Use group G and create graph G in relation with WordNet

4: Mark peak score in graph G

5: Allocate sense with peak score

6: If there is no sense in graph G, allocate the original sense

7: Generate word sense

Fig. 1. WSD Algorithm based Collocations

4 Experiments and Results

This research used the NLTK (the Natural Language Toolkit), which is a strong

natural language processing package of Python, to analyze the morpheme and process

collocations. The data used in the experiment was conducted within data sets of

English words in the Senseval-2, which analyzes the recent functions of the WSD

system. When using the contextual words, which appears within the context

surrounding ambiguous words, the window size of the context has to be considered.

In this paper, considering the size of data sets, Windows 5 was selected as the default

value.

Before constructing graph, if the collocation in the sentence is comprehended, it

increases the connectivity measurement value. The WSD method using collocation

can be considered as a good method regarding its processing speed and performance,

if it is possible to extract sufficient amount of collocations. However, if the

collocation rarely appears in literature, it is difficult to use WSD. Therefore,

considering two dimensions, efficiency and performance; the word sense

disambiguation was made using collocations, and for the words that could not use

collocations, graph centrality algorithm was used for word sense disambiguation. By

using this mixed method, this research was able to achieve good experimental results.


Copyright © 2015 SERSC 79

In Table 2, the precision of the methods suggested in this paper and that of the

contrasting graph-based methods are compared.

Table 2. Comparison with related work

[5] [6] This

paper

Senseval-2 precision 59.54 62.2 66.5

5 Conclusion

In this paper, a new unsupervised WSD method was suggested that does not require

tagged Corpus and shows better performance than the existing unsupervised WSD

method. The algorithm of this paper increased efficiency and constructed graphs by

using collocations to measure the best graph connectivity. Also, it measured the

connectivity using degree centrality algorithms. Moreover, the assessment results

imply that the method of measurement suggested in this paper, which is based on

graph using collocations, yield a more accurate graph connectivity measurement value

than the conventional methods.

References

1. Cho J. K., Shin K. C.: A Graph-based Word Sense Disambiguation using Measures of Graph

Connectivity. In: KIIT, Vol. 12, No. 6, pp. 143-152 (2014)

2. Sinha R., Mihalcea R.: Unsupervised Graph-based Word Sense Disambiguation Using

Measures of Word Semantic Similarity. In: In IEEE International Conference on Semantic

Computing (ICSC) (2007)

3. Navigli R., Lapata M.: An Experimental Study of Graph Connectivity for Unsupervised

Word Sense Disambiguation. In: Pattern Analysis and Machine Intelligence, IEEE

Transactions on, IEEE Computer Society (2010)

4. Lewis M.: Language in the Lexical approach. In: M. Lewis (Ed.), Teaching collocation:

Further developments in lexical approach, pp. 126-153 (2000)

5. Agirre E., Soroa A.: Personalizing pagerank for word sense disambiguation. In: Proc. of

EACL, pp.33-41 (2009)

6. Hessami E., Mahmoudi F., Jadidinejad H.: Unsupervised Graph-based Word Sense

Disambiguation Using Lexical relation. In: International Journal of Computer Issues (IJCSI),

Vol. 8, Issue 6, No 3, pp. 225-230 (2011)


80 Copyright © 2015 SERSC

a graph-based word sense disambiguation using...

Documents